Datadog LLM Observability Pricing: $14,475 a Month at 100M Spans vs Seat Plans

Datadog LLM Observability costs $14,475 a month at 100 million spans in a 10-span agent model, on annual list pricing. Braintrust Pro lands at $3,159 for the same run at small payloads. Datadog bills only model calls, so the per-span label hides a meter that moves with your agent's design. Seat plans hide something else: a usage rate you cannot read.

Datadog LLM Observability pricing at 1M, 10M and 100M spans

Datadog LLM Observability pricing starts at $160 a month on annual billing for 100,000 LLM spans, then $3.50 per 10,000. For an agent emitting 10 spans per run, 4 of them model calls, that is about $269 a month at 1 million total spans, $1,560 at 10 million and $14,475 at 100 million.

Datadog now sells the product as Agent Observability, and buyers still search for the older name. I use the older name where it matches the query and the newer one where it matches the price list. An earlier post set out why LangSmith, Braintrust and Arize bill three different units. This one puts Datadog beside them and does the arithmetic.

Totals are Zan Digital calculations from published list rates. Model inference for judge calls, negotiated discounts, Enterprise plans and taxes are excluded. Every input is in the next section.

The workload behind the LLM observability cost table

A cost table without its inputs is a marketing asset. These are the inputs, so you can replace any of them with your own and recompute the row.

The span mix and payload sizes are illustrative, not measured from a production system. Replace them with one instrumented run of your own agent.

What Datadog LLM Observability pricing actually meters

The phrase "per span" misleads anyone arriving from a tracing tool. Datadog's product page says the LLM span is the only span it bills on, and that tool, workflow, agent, embedding and retrieval spans are all free. Your agent can add retries around a tool without adding a dollar, and each extra model call adds cost immediately.

I think this is the most buyer-friendly meter of the four, and the one most likely to be misread in a budget. A team that multiplies its OpenTelemetry span count by the list rate overstates the bill by 2.4 times in this model.

Span mix moves the bill more than the volume tier

Datadog's cost is close to linear above the first 100,000 LLM spans, so the share of model calls is the variable that matters. Hold the model at 100 million total spans and 1 million judge calls, then change only the agent's mix.

At 20% LLM spans, annual billing costs $7,475 a month. At 40% it costs $14,475. At 60% it costs $21,475. That is the same traffic and the same vendor, with a spread of 2.9 times from agent design alone.

Evaluators sit inside this meter as well. Datadog lists unlimited evals on both plans, but says LLM calls made during an eval run count as LLM spans. Sampling is therefore a cost lever as much as a statistical choice, and the design of your eval suite belongs in the budget conversation.

Commitment term is worth 30% at list

The annual rate is $3.50 per 10,000 LLM spans against $5 on-demand. At 100 million spans that gap is $6,215 a month, which makes $14,475 roughly 30% below the on-demand figure. Datadog says multi-year and volume discounts exist, and it publishes no thresholds for them.

I would not sign annual on a forecast from a pilot. A 12-month commitment is a bet on your span mix, and agent architectures change faster than procurement cycles. Before you commit, read the concessions worth asking for on usage-based contracts.

LangSmith pricing: a seat fee plus a trace rate you must request

LangChain's pricing page lists Plus at $39 per seat per month with up to 10,000 base traces, then pay as you go. LangChain defines a trace as a single execution that can include many steps. So 100 million spans is only 10 million traces here, and span count disappears from the invoice.

That is a genuinely different bargain. A long agent run costs the same trace as a short one, up to a limit of 25,000 runs per trace. The difficulty is that the text of the pricing page and billing docs I could read on 8 November 2026 shows no per-trace figure. The billing docs point back to the pricing page for it.

Third-party trackers quote rates, and they disagree with each other. None could be confirmed against a LangChain page, so none appears here. The test rates below are round numbers chosen to show the shape of the bill. They are not quotes.

The last row is the useful one. If the rate LangChain quotes you sits below it, LangSmith is cheaper than Datadog annual list pricing for this workload. Extended retention upgrades are excluded.

Seats stop being the main cost almost immediately

Seat pricing punishes broad access, and it does so early. At 1 million spans, Datadog annual costs $268.50, while 7 LangSmith seats cost $273 before a single trace charge. Any team wider than 6 people pays more for access alone. The wider dynamics are covered in the analysis of seat compression in SaaS pricing.

Above 1 million spans the relationship flips, because the fixed $195 becomes a rounding error beside whatever the trace rate turns out to be.

The event cap arrives before the invoice does

Plus carries a limit of 500,000 trace events an hour, and an event is the creation or update of a run. At 100 million spans over 720 hours you average 138,889 runs an hour. If each run produces one create and one update, that is about 277,778 events, or 56% of the cap on average.

A peak of 1.8 times average traffic reaches the ceiling. So the 100 million column is really an Enterprise conversation with custom pricing, and you should read it that way.

Braintrust pricing: where a byte meter beats a span meter

Braintrust Pro costs $249 a month with unlimited users, 5 GB of processed data and 50,000 scores. Overage runs at $3 per GB and $1.50 per 1,000 scores. Its plans page defines processed data as total bytes ingested across logs, experiments and datasets, measured at ingestion, so deleting data later does not reduce the charge.

This is why the Braintrust line grows only 12.7 times from 1 million to 100 million spans while Datadog's grows 53.9 times. The fixed fee absorbs more of the small workload, and the unit price is low. At high volume with modest payloads, I would start here.

Payload size is the assumption to measure first

Braintrust states no average span size, so my 5 KB is an assumption. Agents that log full retrieved documents or long conversation histories will run far larger. At 20 KB per span, the 100 million span bill rises to $7,659, and Datadog's multiple falls from 4.6 times to 1.9 times.

The tipping point is about 42 KB per span at 10 million and 100 million spans. At 1 million spans it is only about 11.5 KB, because the fixed fees are so close together. Above those sizes, Datadog's indifference to payload wins.

Scores behave like a second meter

At 100 million spans, scores contribute $1,425 of the $3,159 total, which is 45%. Doubling the sampling rate to 20% adds another $1,500 a month. It will not move the Datadog meter by anything like the same proportion unless those scores run through LLM judges, in which case it moves that one too.

Arize AX pricing stops at the included tier

Arize's pricing page lists AX Pro at $50 a month with 50,000 spans, 10 GB of ingestion, 30-day retention and unlimited users. Its span row reads "Included 50k/month" and gives no overage figure. The docs I could open did not define a billable span or say what happens above the allowance.

All three volumes in this model exceed that allowance, so a total would be invented. What can be computed is the rate at which Arize ties Datadog annual pricing. That rate is about $230 per million spans at 1 million total, $152 at 10 million and $144 at 100 million, before any storage charge.

The rate you need to request therefore has two parts, spans and gigabytes. At 5 KB per span, the 10 GB allowance is gone at about 2 million spans. Note also that Dynatrace announced a definitive agreement to acquire Arize on 13 August 2026, so ask how long any quoted rate will hold.

Retention changes the LLM observability cost gap

The main cost table is not perfectly like for like. Datadog's base is 15 days and LangSmith's is 14, while Braintrust Pro and Arize AX Pro include 30. If your incident reviews need a month of history, Datadog's line has to move.

The 30-day row is $5 per 10,000 LLM spans on annual billing. Read as an all-in rate, the 100 million span workload costs $20,610 a month, which is 6.5 times Braintrust at 5 KB. Datadog's FAQ calls these "add-ons", though, and read as an addition to the $3.50 base the same bill is $34,925.

I lean to the all-in reading, because each row is priced as a complete retention tier. Datadog's pages do not settle it, and a $14,315 monthly swing is worth one email to your account team. For why that window matters operationally, see how agent drift shows up in production.

Where this cost model is weakest

A model that looks this precise deserves suspicion, so here are the four places it could mislead you.

The workload is mine, not yours

The 10-span shape, the 40% LLM share and the 5 KB payload are illustrative. No published benchmark supports them, and the conclusion can change. A chat application with one model call per trace would look nothing like this, and neither would a planner that fans out to 50 tools.

List price is the wrong price at 100 million spans

Nobody spending $14,475 a month on telemetry should pay list. Datadog advertises undisclosed volume discounts, and all four vendors route large accounts to custom contracts. Treat the right-hand column as an anchor for negotiation, not a forecast. Teams already on Datadog may also have allotments in existing agreements, and Datadog's exposure to AI-native customers gives both sides reasons to negotiate.

The strongest case against my ranking

Cost is not the deciding variable for many buyers. A team already running Datadog APM gets its model calls in the same place as its infrastructure traces, and that consolidation may be worth a large premium over a cheaper standalone tool. Datadog is also the only one of the four whose variable rate I could read in full, which is a real advantage for anyone building a budget.

Frequently asked questions

Where to start

Instrument one real agent run this week and write down 3 numbers: total spans, LLM spans and average payload in kilobytes. Those replace every assumption in the table above. The exercise takes an afternoon, and it will tell you more than a vendor demo.

Then send LangChain and Arize one identical request for a written usage rate at your 12-month volume, including retention and storage. A vendor that will not put the rate in writing has told you something about the renewal. Pair the answer with how vendor meters price a single agent task so the telemetry bill sits beside the inference bill it is meant to explain.

Related reading